Repository navigation
[PR] test(preflight): cover calibrated exhaustion on the DSH Responses lane - #3968
Conversation
(ai: qwen/qwen3.8-27b) 查重: 不重复。 检索 open+closed 的 issue/PR(关键词 calibrated futility / responses lane / preflight + 失败签名 标签: pr(既有)+ dsh(DSH 插件车道)+ scope/ci(纯回归测试补强)+ p3(测试覆盖缺口,无用户可见行为变更)。 状态: PR 为 draft、head |
(ai: qwen/qwen3.8-27b) PR 审核结论(独立重审,非复读 issue) 已完整走查 diff(+95/-5,仅 正向验证:把该 commit cherry-pick 到当前 master( 负向验证(测试有牙):在 rebased head 上精确回退 #3650 的修复 hunk(preflight.ts 的校准比较全部还原 + server.ts fail-fast 引用还原为 即引用了未校准的 raw 尺寸(≈18160 < 窗口 30000)——与作者声明的 v0.1.202 失败模式一字不差。另有一个值得写进记录的事实:pre-fix 代码下本 case 的折叠并未被跳过(summaries≥1 仍过),说明这个 Responses case 钉的是「fail-fast 引用尺度」这一半(server.ts 侧),与原 chat case 钉的「futility gate 误判」一半互补,两条腿各管一个回归面。 八问
代码质量:命名达意(responsesSse/responsesPlain/responsesBashPair/postResponses/MockWire)、风格与文件既有惯例一致、无死代码、无新依赖;fixture 的事件序列与 OpenAI Responses SSE 实际形状吻合。达标。 唯一程序性问题:分支基线 判定:审核通过,待 rebase + un-draft 后可合并。 |
📦 Merged Build PublishedThis merge to master was published as npm install -g billion-context@0.0.691-master.5d2ea3c # this exact build
npm install -g billion-context@master # always the newest merged build
|
The calibrated preflight regression covers Chat Completions, while DSH's plugin lane can send Responses requests. This adds a Responses SSE fixture and a
x-bili-plugin: dshcase that exercises actual summary folding, then checks that calibrated exhaustion refuses the main request and reports a context figure above the tested input budget.The runtime fixes are already merged in #3650 and #3649. This PR adds the missing wire/lane regression; it does not resubmit those changes or alter overflow policy. Refs #3645.
Validation:
tests/issue3645-calibrated-futility.test.tspass against the corrected sources.440c5114and on this test-only head reports the same existingTS2307error formysql2/promiseinsrc/storage/mysql.ts; this PR leaves that unrelated dependency change untouched. The four targeted cases pass on the current base plus this change.The test uses a loopback fake upstream and synthetic input. It requires no credentials, real model request or private diagnostic material.